Overview

This is the code for running the analyses in The Supervision Horizon: An Exploration of the Mechanics of On-Policy Self-Distillation. There are three core analyses:

1. Computing core metrics across the sequence
2. Measuring the gradient norm from training on the first vs last 25% of the sequence
3. Measuring the within-problem cosine similarity of gradients from training on the first vs last 25% of the sequence

Getting Started

This project uses uv. Collect the dependencies with:

uv sync

To run each experiment, do:

uv run -m analysis.experiment_1_core_metrics

uv run -m analysis.experiment_2_gradient_norm

uv run -m analysis.experiment_3_gradient_similarity

Each experiment has the following arguments you can adjust:

• model-name
• num-problems
• num-rollouts
• temperature
• max-new-tokens
• top-k
• output-dir